I deeply love this idea of specialized LLMs for search. It's also extremely confusing to me how rough Google's entrance here is.
When I, a human, need an answer to anything moderately complex, it's unlikely that I get it on the first (pre-AI) round of google searching. Simple stuff, sure, but more likely I'll need to go 2-5 rounds. Maybe click a few links. Double-check my assumptions.
An LLM that can do that quickly seems like a slam dunk. I wonder what other problems benefit from that 10x-100x increase in context + 2-5 rounds with the LLM.
About 15 years ago I would sometimes spend hours on Google image search discovering childhood toys and filling in vague memories of locations or things. I tried this recently and it’s basically impossible. I actually get to the end of the search results in like 3 minutes and the quality is horrible now.
I think in a lot of ways Google peaked and is now on the decline into a profitable but much less relevant services company.
Maybe the web is really that much worse, and the SEO tactics so hostile to genuine content.
But, honestly...just bring back the old google, where I have all kinds of search modifiers to perform exactly the search I want, that just returns all the matching results that have been indexed. Let me sort out the rest. How do you remove the ability to do "exact text search"? It's the most basic of search functions. Remember having "|" modifiers? AND modifiers?
If google launched that again, even as a separate engine, I think they'd have a good product again. Maybe being good doesn't pay the bills for google, though.
Recently learned I can just add a question mark to the end of my Kagi search to get an assistant answer. Kagi was already great at surfacing the most relevant pages, but now I often don't even need to click.
For example, a couple of days ago I described a problem with my refrigerator's water dispenser to Google Gemini, and it told me exactly how to fix it. I then went looking for a video and fixed the thing in under 15 minutes. The only way that Gemini could have been better is if it linked to a video itself.
Do you mean search into less-well-known topics? Or something else?
I’m astonished at how bad search in Gmail is. I searched for “iPhone receipt” to find out when I purchased my current iPhone, and it pulled up every single email I’ve ever gotten from Best Buy, since they all have a link in there to buy an iPhone, and they all have a link in there to look up a receipt from them.
I know that those emails have both the words “iPhone” and “receipt” but I feel like The Search Engine Company That Also Does Email should have a smarter search engine in their email.
I’m pleasantly surprised that I can ask Gemini for help with it if I specify “look through my Gmail for this information”, but it takes it a minute or two to find it.
I know it can be a deep time sink, but I notice more and more how much deeper my understanding is of a certain problem/best-practice after developing the neuropathways involved in crawling between reddit, stack overflow, etc, to get to the proper solution. I love the instant answer from google ai, but I also notice an itch to purposefully force myself to ignore it when time allows.
Looks good as I use something similar with the SearXNG MCP, but a shame this isn't an open weight model. There are some wrappers around SearXNG which seem to reduce the token counts returned thus making it easier for the calling model to understand, but a full dedicated model for search is nice. How does it compare with Perplexity, Gemini with search, and Parallel AI? Those are the cloud providers of search based models that I've seen so far.
hey its a search agent for YOUR own data but can also work over the web. Mixedbread is focusing on providing evidence for agents for your internal data. Toast can interact with any search api. You should be able to provide the SearXNG api to it and it should be good to go. Here the default harness: https://github.com/mixedbread-ai/toast-harness
I'm pretty confident we could steal The Though Emporium's Single use thermite based instant hot dog [1] design and make instant packaged toast, fresh to order
I guess someone who has used a search agent (or a dedicated subagent) can speak when I'd reach for a tool like this vs either just 1) a smaller general model or 2) a non-llm approach to the problem? Like it's interesting I'm just curious how a search agent compares to say a model with dedicated rag pipelines is that much different?
the issue with smaller general models (see at the charts) are way behind the frontier models when it comes to search. we've found that there is huge uplift of having a fast dedicated model. from our perspective, having a very good index is the biggest lever and then having a specialised model.
A good index is a software and LLM problem if using the LLM for indexing. Are you looping "agents" in an embedding and encoding cycle before retrieval? There are thousands of RAG agents at this point and RAG is still not super great. A dedicated specialized model? You want to take on Qwen3.6 or Qwen3.8 wrapped a pi.dev harness agent that has been dedicated to be the "search" agent? How would you stack up?
you can look it up in the blog. RAG is not super great because of two reasons, single embedding vector models are not that good and stopped improving and second most models are not good at looking up information. we spend great time on improving the modeling side by inventing on the indexing level [1, 2]. and now we trained our model to be very good at search. it is matching the quality of Opus 5 and GPT 5.6 Sol while being faster. it helps your main agent to do the task at greater quality, while reducing cost per task.
Mixedbread Search is a multimodal & multilingual search product, where you can upload any kind of data and make it searchable. Its powered by Wholembed [1] v3, a late interaction retrieval model.
I don't think so, but could be wrong. It seems to be a specialized layer like a lora or a merged model. So a RAG-agent-model thingy. No clue if it actually works. I don't understand why I would want to use it.
the thing is most agents waste most of their tokens looking up information which can cause context rot. most small models are not as good as looking up information. this model helps to lookup information for your main agent, which helps you to save tokens and still maintain quality.
Long time user of your embedding models. I'm trying to understand how this works and if I can leverage it.
It sounds like this is a new layer on top of your existing storage layer? So to use this, would I need to give you all of my data first? Or is there a version that can be run on prem?
yes for the retrieval benchmarks. For officeqa pro v2 we used Codex (as databricks did) and for Harvey LAB we used the vanilla harvey benchmark. For these benchmarks we added minimal tools to use mixedbread search and toast 1.
Looking through your blog it is very much keeping the brand baked in. It would be a nice little nod to toss a couple sentences about the branding on the about page or somewhere so if someone wants to know how you came to it they can. Leaving it to "wink, nod, inside joke you'll never know" is a bit off-putting when you are building a brand around the theme. It also makes it more memorable for those that read the blurb.
If you get big, the naming no longer matters. We made jokes about the Wii until everybody had one. But if you don't get big, and most people have no idea what they're looking at when they see your product for the first time, naming definitely matters.
Me, looking at an enterprise product: “Hmm, but do they guarantee FULL lore? I don’t really need to know what the lore is, just the assurance that it is full”
I know everyone loves to hate on google but i find search overviews and asking gemini to search for things way faster than any alternative. I was curious about a development near me and asked literally that and gemini pulled court records in about 20 seconds
Anyway, back to this - it seems to be more like the AI equivalent of algolia than google
I deeply love this idea of specialized LLMs for search. It's also extremely confusing to me how rough Google's entrance here is.
When I, a human, need an answer to anything moderately complex, it's unlikely that I get it on the first (pre-AI) round of google searching. Simple stuff, sure, but more likely I'll need to go 2-5 rounds. Maybe click a few links. Double-check my assumptions.
An LLM that can do that quickly seems like a slam dunk. I wonder what other problems benefit from that 10x-100x increase in context + 2-5 rounds with the LLM.
About 15 years ago I would sometimes spend hours on Google image search discovering childhood toys and filling in vague memories of locations or things. I tried this recently and it’s basically impossible. I actually get to the end of the search results in like 3 minutes and the quality is horrible now.
I think in a lot of ways Google peaked and is now on the decline into a profitable but much less relevant services company.
Google search is so bad compared to peak google.
Maybe the web is really that much worse, and the SEO tactics so hostile to genuine content.
But, honestly...just bring back the old google, where I have all kinds of search modifiers to perform exactly the search I want, that just returns all the matching results that have been indexed. Let me sort out the rest. How do you remove the ability to do "exact text search"? It's the most basic of search functions. Remember having "|" modifiers? AND modifiers?
If google launched that again, even as a separate engine, I think they'd have a good product again. Maybe being good doesn't pay the bills for google, though.
Kagi
Recently learned I can just add a question mark to the end of my Kagi search to get an assistant answer. Kagi was already great at surfacing the most relevant pages, but now I often don't even need to click.
Yandex image search is a lot better now.
What kind of search do you mean?
For example, a couple of days ago I described a problem with my refrigerator's water dispenser to Google Gemini, and it told me exactly how to fix it. I then went looking for a video and fixed the thing in under 15 minutes. The only way that Gemini could have been better is if it linked to a video itself.
Do you mean search into less-well-known topics? Or something else?
Have always joked with colleagues about how hard it is to find content in Google Docs, the office suite built by a search company.
I’m astonished at how bad search in Gmail is. I searched for “iPhone receipt” to find out when I purchased my current iPhone, and it pulled up every single email I’ve ever gotten from Best Buy, since they all have a link in there to buy an iPhone, and they all have a link in there to look up a receipt from them.
I know that those emails have both the words “iPhone” and “receipt” but I feel like The Search Engine Company That Also Does Email should have a smarter search engine in their email.
I’m pleasantly surprised that I can ask Gemini for help with it if I specify “look through my Gmail for this information”, but it takes it a minute or two to find it.
I know it can be a deep time sink, but I notice more and more how much deeper my understanding is of a certain problem/best-practice after developing the neuropathways involved in crawling between reddit, stack overflow, etc, to get to the proper solution. I love the instant answer from google ai, but I also notice an itch to purposefully force myself to ignore it when time allows.
ime google ai is wrong often enough that I'm still doing that to verify its results
At least validation seems faster than without, but you get what you pay for when it comes to llm intelligence
Gemini search isn’t great by default, but Deep Research gives noticeably better results.
I've been using https://www.voyageai.com/ for this in production. Might try this out to see how it compares.
Looks good as I use something similar with the SearXNG MCP, but a shame this isn't an open weight model. There are some wrappers around SearXNG which seem to reduce the token counts returned thus making it easier for the calling model to understand, but a full dedicated model for search is nice. How does it compare with Perplexity, Gemini with search, and Parallel AI? Those are the cloud providers of search based models that I've seen so far.
hey its a search agent for YOUR own data but can also work over the web. Mixedbread is focusing on providing evidence for agents for your internal data. Toast can interact with any search api. You should be able to provide the SearXNG api to it and it should be good to go. Here the default harness: https://github.com/mixedbread-ai/toast-harness
I really wanted this to be a hardware startup - the Juicero of toast.
Sadly its another software company.
Selling you a subscription for slices of toast in individual cartridges for their toasting hardware?
I'm pretty confident we could steal The Though Emporium's Single use thermite based instant hot dog [1] design and make instant packaged toast, fresh to order
[1] - https://www.youtube.com/watch?v=8X1_HEJk2Hw
"The perfect slice, every time. No muss no fuss."
Seems to me if it doesn't also apply uniform melted butter it would be a hard sell.
I guess someone who has used a search agent (or a dedicated subagent) can speak when I'd reach for a tool like this vs either just 1) a smaller general model or 2) a non-llm approach to the problem? Like it's interesting I'm just curious how a search agent compares to say a model with dedicated rag pipelines is that much different?
the issue with smaller general models (see at the charts) are way behind the frontier models when it comes to search. we've found that there is huge uplift of having a fast dedicated model. from our perspective, having a very good index is the biggest lever and then having a specialised model.
A good index is a software and LLM problem if using the LLM for indexing. Are you looping "agents" in an embedding and encoding cycle before retrieval? There are thousands of RAG agents at this point and RAG is still not super great. A dedicated specialized model? You want to take on Qwen3.6 or Qwen3.8 wrapped a pi.dev harness agent that has been dedicated to be the "search" agent? How would you stack up?
you can look it up in the blog. RAG is not super great because of two reasons, single embedding vector models are not that good and stopped improving and second most models are not good at looking up information. we spend great time on improving the modeling side by inventing on the indexing level [1, 2]. and now we trained our model to be very good at search. it is matching the quality of Opus 5 and GPT 5.6 Sol while being faster. it helps your main agent to do the task at greater quality, while reducing cost per task.
[1]: https://www.mixedbread.com/blog/multimodal-late-interaction-... [2]: https://www.mixedbread.com/blog/wholembed-v3
interesting, thanks for the reply.
Article should probably explain what "Mixedbread Search" is.
Mixedbread Search is a multimodal & multilingual search product, where you can upload any kind of data and make it searchable. Its powered by Wholembed [1] v3, a late interaction retrieval model.
[1]: https://www.mixedbread.com/blog/wholembed-v3
so, it works like notebooklm but without relying on RAG?
I don't think so, but could be wrong. It seems to be a specialized layer like a lora or a merged model. So a RAG-agent-model thingy. No clue if it actually works. I don't understand why I would want to use it.
you can try it here: https://dwarkesh-search-demo.vercel.app/
the thing is most agents waste most of their tokens looking up information which can cause context rot. most small models are not as good as looking up information. this model helps to lookup information for your main agent, which helps you to save tokens and still maintain quality.
Long time user of your embedding models. I'm trying to understand how this works and if I can leverage it.
It sounds like this is a new layer on top of your existing storage layer? So to use this, would I need to give you all of my data first? Or is there a version that can be run on prem?
you can plugin your existing stack and use it via an openai compatible client.
https://www.mixedbread.com/docs/agent/chat-completions
Are the benchmarks comparing just the models while keeping the harness the same (the open-source Toast harness)?
yes for the retrieval benchmarks. For officeqa pro v2 we used Codex (as databricks did) and for Harvey LAB we used the vanilla harvey benchmark. For these benchmarks we added minimal tools to use mixedbread search and toast 1.
> performs best with Mixedbread Search, but it can work with any search backend
Bread-first search, is it?
What are these names? before opening I thought it was going to be about bread
Nice, a model for google-fu.
Some days I have no idea what the fuck I am looking at.
Bah, i was looking forward to a new toaster!
I was looking forward to a new version of the macOS CD/DVD burning software.
I love to spend a dollar for a 70% correct search result
Dunno about this branding/naming scheme - every time it comes up we have to double-check that it isn't some spoof/joke page
there is full lore around the naming. i can guarantee you that we are pretty dedicated around our research and product.
Looking through your blog it is very much keeping the brand baked in. It would be a nice little nod to toss a couple sentences about the branding on the about page or somewhere so if someone wants to know how you came to it they can. Leaving it to "wink, nod, inside joke you'll never know" is a bit off-putting when you are building a brand around the theme. It also makes it more memorable for those that read the blurb.
An inside joke that confuses your target audience and makes you sound like a joke, is probably not what you want as your brand.
The brand only needs to last until the Google acquihire.
(FWIW I like the branding. Hugging Face doesn't seem to struggle because of its name either)
If you get big, the naming no longer matters. We made jokes about the Wii until everybody had one. But if you don't get big, and most people have no idea what they're looking at when they see your product for the first time, naming definitely matters.
lore does not make it a good or professional choice
Me, looking at an enterprise product: “Hmm, but do they guarantee FULL lore? I don’t really need to know what the lore is, just the assurance that it is full”
All the stupid made up names are taken, startups have to take regular words and make them meaningless now
I’m disappointed this is some AI thing and not a breadboard company.
I'm disappointed it doesn't burn CDs.
I know everyone loves to hate on google but i find search overviews and asking gemini to search for things way faster than any alternative. I was curious about a development near me and asked literally that and gemini pulled court records in about 20 seconds
Anyway, back to this - it seems to be more like the AI equivalent of algolia than google
Damn, now I'm hungry. All this bread talk.